Deep learning based high performance classification architecture for low-altitude aerial images

One of the major factors creating advance prospects in the aerial imaging classification solutions market is the recently published drone policies by Government of India and availability of artificial intelligence-based technologies. The images in low-altitude aerial datasets are inherently different from standard datasets in terms of the appearance cues and the number of bounding box hypotheses. The appearance cues exist due to the present challenges in low-altitude aerial images such as change in viewpoints, arbitrarily orientation and occluded objects. The wide coverage of objects in low-altitude aerial images accounts for a large number of objects in aerial images resulting in complex and multiple bounding boxes. These challenges trigger a need for powerful classification architectures for low-altitude aerial images. This research paper discusses high-performance classification technique based on powerful feature extractor proposed for low-altitude aerial images. The proposed classification architecture makes use of the new improved VGG16 network and dilated ResNet50 model in which fusion takes place between various transformed feature maps. The fusion helps in embedding extra semantic information which further aids in accurate classification of low-altitude aerial images. The performance evaluation is done on approximately 23 k images with different classes of objects gathered from various benchmark low-altitude aerial datasets. The proposed classification architecture achieved a validation accuracy of 99.70% and test-set accuracy of 96.23% which is better than other classification models.